Looks sensible to me. Minor observations:

1. It will blow up with a division by zero if z = 0, however, that's probably the correct behaviour.
2. The two checks for overflow are clever - if we overflow, then we have definitely exceed our invariant, and so need to do the reduce step.
3. I've suggested recording one extra invariant above.
4. I've run a few different versions of this test code, and it has always produced the correct answers, so I think it is working fine.

Will have a look today.

This looks good to me - including a random hashseed seems like a great idea. I guess someday it might be neat to hash the fport and lport values, rather than xor them, but that should wait for another day if this has already been running in Netfilx.

Also looks good to me. Slightly pedantic suggestions:

Hi Bapt,

Great - I'll commit it in the next day or two, and ask for permission to MFC.